Agentic AI for AdaptFi
A live, visual explanation of how LangGraph turns AI from a one-shot chatbot into a controlled expert workflow for adaptation finance screening.
Why a chatbot is not enough
A banker or consultant does not need a clever answer. They need a repeatable process, source evidence, human review, and an audit-ready result.
Normal AI chatbot
Agentic AdaptFi workflow
The work is broken into accountable steps. Each step creates evidence. The human expert reviews. The final result is locked into an audit record.
LangGraph in plain English
LangGraph is easiest to explain as a map of workstations. A project file moves through the map. Some stations are AI, some are code, and some are human review gates.
State
The shared project file. It carries facts, draft answers, evidence, reviewer edits, and the final status.
Node
A workstation in the process. One node may extract fields, another may draft eligibility answers, another may run deterministic checks.
Edge
The route between stations. It decides what happens next, such as move to review, ask for correction, or commit the result.
Interrupt
A deliberate human pause. The workflow stops and waits until the expert reviews, edits, approves, or sends back.
Checkpoint
A saved position in the workflow. AdaptFi can pause, resume, survive failure, and continue later without losing the screening history.
The AdaptFi agentic workflow
Each uploaded project becomes a screening thread. The thread moves through specialist agents, deterministic checks, a human review interrupt, and a final signed record.
Screening state
The current working file of one project.
{
status: "queued",
district: "pending",
hazards: [],
measures: [],
review: "not started"
}
Message to remember
Agents do the preparation. Code computes the formal rollup. Humans sign the determination of record.
Audit trail builds as work happens
Each step leaves a trace.
For non-technical users
Think of this as an expert assembly line with memory, evidence, pauses, controls, and a final signed output.
AI drafts. Human experts sign.
This is the most important trust principle in AdaptFi. The AI prepares evidence-backed drafts. The expert remains responsible for the determination.
Agent draft
Expert review gate
The reviewer checks each drafted answer, reads the evidence, and decides what becomes part of the record.
Signed determination
Only the confirmed answers are used to produce the formal AdaptFi output.
SIGNED
Where AI stops and code takes over
This distinction makes the platform defensible. AI drafts the reasoning. Deterministic code computes the formal rollup after human confirmation.
AI drafting layer
Useful for judgement-heavy preparation.
- ✓Extract facts from uploaded project documents.
- ✓Suggest relevant adaptation measures from the knowledge base.
- ✓Draft eligibility and DNSH answers with citations.
- ✓Assemble readable narrative sections for review.
Deterministic control layer
Used for final formal outcomes.
- ✓Compute MDB/IDFC eligibility rollup from confirmed answers.
- ✓Flag DNSH or maladaptation outcomes from confirmed responses.
- ✓Pin methodology version, hazard snapshot and inputs hash.
- ✓Persist final output artefacts and audit events.
The audit vault
AdaptFi is useful because a future reviewer can understand not only the conclusion, but exactly how the conclusion was produced.
Agents earn trust over time
The system starts conservative. Human experts review everything. Their corrections become the training signal for better future drafts.
The one-slide summary
AdaptFi uses agentic AI to convert adaptation finance screening from a manual, consultant-heavy process into a controlled, evidence-backed, human-signed workflow.